[feature-audit] plot_features degenerates to (i,t) with no plot_id level — Niger, Mali, Malawi, GhanaLSS (split from #513)

[ligon]

Split from #513. The deep-dive found a SECOND, distinct duplicate-collapse class that was being conflated with #513's (t,i,plot_id) collapse.

Verified

Feature('plot_features')([C]) emits Canonical index over ['i','t'] had N duplicate tuple(s); collapsed via groupby().first() — i.e. the index has no plot_id level at all:

country	(i,t) dups
Malawi	60 (in addition to its 23 on (t,i,plot_id))
Niger	59
Mali	~8
GhanaLSS	2

Distinct from #513

#513 is the (t,i,plot_id) collapse (a real plot index colliding under id_walk). This class has no plot dimension — plot_features for these countries is effectively household-grain, so the canonical (t,v,i,plot_id) degenerates to (i,t) and multiple plots per household collapse via first().

Next step (per-country)

Either populate a real plot_id from the source (if one exists) or document plot_features as HH-grain for these countries and reconcile the canonical index expectation. Relates to #469 (EthiopiaRHS plot-grain blocked).

Severity: B/C (silent loss vs modeling gap, varies by country). Found via the backlog-workflow deep-dive on #513.

[ligon]

Adversarial diagnosis (workflow + red-team) — re-scope: this is mostly a sample-table collapse, not plot_features

Both red-team skeptics monkeypatched _normalize_dataframe_index (the warning emit site) to record which table actually triggers each (i,t) collapse. Result:

country	sample-table dups	plot_features (t,i,plot_id) dups
Niger	59 (identical → benign)	0
GhanaLSS	2 (identical → benign)	0
Mali	8 (divergent weight/panel_weight → real loss)	6 (divergent Area/SoilType → real loss)
Malawi	60 (divergent panel_weight/strata/Rural/District → real loss)	23 (divergent → real loss)

Two corrections to the original framing

1. The (i,t) warnings fire from the sample table, rebuilt inside _join_v_from_sample() for every household-grain table — so the harness mis-attributes them to whatever table triggered the v-join (here, plot_features). The (i,t) symptom is a sample / panel-id issue, not a plot_features one.
2. But it is not "zero plots collapse / not a plot_features bug" either: Mali (6) and Malawi (23) genuinely drop distinct plots from plot_features itself on (t,i,plot_id) — those belong to [feature-audit] plot_features / all: Canonical index over ['t', 'i', 'plot_id'] had 163 duplicate tuple(s); collapsed via groupby().first(), dropping those rows (possible sil... #513 (now folded in there).

Disposition: re-scope #536 → sample-table (i,t) collapse

• Niger / GhanaLSS: benign (identical rows) — at most a reporting fix so _join_v_from_sample doesn't blame the caller's table for a sample-internal collapse.
• Mali (8) / Malawi (60): real sample-table silent loss — two distinct same-wave households share a baseline id and first() drops a genuinely-different sample row (different weight / cluster v / district). Same family as Panel-ID transitive-chain _N-suffix collision in update_id collapses distinct Malawi 2019-20 households #504 (panel-ID transitive _N-suffix collision).
• The plot_features (t,i,plot_id) pieces (Mali 6, Malawi 23) → [feature-audit] plot_features / all: Canonical index over ['t', 'i', 'plot_id'] had 163 duplicate tuple(s); collapsed via groupby().first(), dropping those rows (possible sil... #513.

No code changed (diagnosis only).

[ligon]

Resolved by #546, merged to master in #555. (Auto-close didn't fire: the fix PR targeted development (not the default branch), and commits use fix(#N): scope style rather than GitHub closing-keyword syntax.)